In July 2019, The Guardian published a story that changed how millions of people thought about voice assistants. Apple contractors - real human beings, sitting in offices - were listening to recordings of Siri interactions. Some of those recordings included sensitive medical conversations, drug deals, and couples being intimate. Apple had not disclosed this practice to users.
The backlash was swift. Apple apologised, suspended the programme, and promised changes. But the deeper question - what exactly does Apple Dictation send to Apple servers, and when does it stop? - never received a clean, complete answer.
If you use Dictation on your Mac, iPhone, or iPad, your voice data may still be leaving your device. Not always, and not in the way it did in 2019. But the distinction between "on-device" and "truly private" is one that Apple has not made easy to understand.
This article breaks down what happens to your voice when you press that microphone button, what Apple retains, how long they keep it, and - most importantly - how to stop it.
Does Apple Dictation Send Data to Apple Servers?
The short answer is: it depends on your device, your operating system version, and your settings. In many cases, yes - your audio or transcription data is transmitted to Apple servers.
Apple introduced on-device dictation processing with iOS 15 and macOS Monterey in late 2021. If you are running a recent version of macOS or iOS on supported hardware, basic dictation can run locally. But "can" is doing a lot of heavy lifting in that sentence.
Several conditions still trigger server-side processing:
- Certain languages are not supported by the on-device model and require cloud processing
- Older hardware without a Neural Engine falls back to server-based transcription
- Siri integration - when dictation is used through Siri rather than the keyboard shortcut - may involve server requests
- The "Improve Siri & Dictation" toggle, if enabled, sends audio samples to Apple for review
Even when dictation runs on-device, Apple's operating system may still transmit metadata, usage patterns, or partial transcriptions depending on which features you have enabled and how the system routes requests.
The gap between Apple's marketing language - which emphasises privacy as a core value - and the actual data flow is where the problem lives. "Processed on-device" does not mean "never leaves the device." It means the transcription model runs locally. Everything else - analytics, crash reports, feature improvement data - operates under a different set of rules.
What Data Does Apple Collect From Dictation?
Understanding what gets collected requires distinguishing between three categories of data: audio, transcription text, and metadata.
Audio Recordings
If you have enabled "Improve Siri & Dictation" in your privacy settings, Apple collects a percentage of your dictation audio. Apple states that these recordings are anonymised - stripped of your Apple ID and associated with a random identifier. After the 2019 controversy, Apple made this feature opt-in rather than opt-out, and now says human review only occurs with your explicit consent.
However, the concept of "anonymised" voice data is itself contested. Research has demonstrated that voice recordings contain biometric markers - pitch, cadence, accent, speech patterns - that can be used to re-identify individuals even when traditional identifiers are removed. Your voice is, by its nature, personally identifiable.
Transcription Data
When dictation is processed server-side, the audio is sent to Apple's servers, transcribed, and the text is returned to your device. Apple's privacy documentation states that audio sent for server-based transcription is not retained beyond the time needed to process the request, unless you have opted into the improvement programme.
But "not retained" and "not logged" are different things. Server-side processing involves routing audio through Apple's infrastructure, where it passes through load balancers, processing nodes, and potentially temporary storage buffers. The distinction between "retained" and "temporarily stored during processing" is a legal one, not a technical one.
Metadata and Usage Patterns
Even with on-device dictation, Apple collects usage analytics: how often you use dictation, how long your sessions last, which apps you dictate into, error rates, and device performance metrics. This data is covered under Apple's general analytics and privacy policies, and while Apple says it is aggregated and anonymised, it is still transmitted from your device.
Apple's privacy practices have improved significantly since 2019, and the company is more transparent than most of its peers. But "more transparent than Google or Amazon" is a low bar when it comes to voice data. The fundamental issue is that any cloud-based voice processing creates a data trail you cannot fully control.
How Long Does Apple Keep Your Voice Data?
Apple's data retention policies for voice data have changed several times since 2019, and the current state is more nuanced than a single number.
For users who have opted into "Improve Siri & Dictation," Apple retains audio samples for up to six months. After that period, a copy may be retained without the random identifier for longer-term machine learning improvement - Apple has previously indicated this secondary retention can extend to two years. Apple does not provide users with a way to access or delete specific audio samples from this pipeline, though you can disable the feature at any time, which Apple says stops future collection.
For standard dictation that is processed server-side (without the improvement toggle), Apple states that audio is deleted "shortly after" processing is complete. The definition of "shortly after" is not specified in public documentation.
For on-device dictation, Apple says no audio leaves the device. This is the cleanest category, but as noted above, it only applies under specific conditions.
| Data Type | Retention Period | User Control |
|---|---|---|
| Audio (improvement programme) | Up to 6 months (+ extended ML retention) | Opt out in Settings |
| Audio (server-side processing) | Deleted "shortly after" processing | Limited - depends on hardware/language |
| Transcription text (server-side) | Not retained after delivery | None |
| Usage metadata | Subject to Apple's general analytics policy | Disable analytics in Settings |
| On-device audio | Never leaves device | Use supported hardware and language |
The practical takeaway: unless you have confirmed that your specific device, operating system version, and language combination supports on-device processing - and you have disabled the improvement programme and general analytics - some form of data is leaving your device when you use Apple Dictation.
How Does On-Device Dictation Actually Work?
Apple's on-device dictation uses a speech recognition model that runs locally on the device's Neural Engine. On Apple Silicon Macs and recent iPhones, this model handles the conversion from audio to text without sending audio to Apple's servers.
The on-device model is genuinely impressive from a technical standpoint. It processes speech in real time, supports auto-punctuation, and handles common dictation commands. For supported configurations, it is a meaningful privacy improvement over the fully cloud-based system that preceded it.
But on-device processing has limitations that Apple does not prominently communicate:
Language coverage is incomplete. Apple's on-device dictation supports a subset of the languages available through server-side processing. If your primary language is not in the on-device group, your audio is sent to the cloud every time you dictate - with no warning or indicator distinguishing the two modes.
There is no visual indicator. When you activate dictation, macOS and iOS do not tell you whether your audio is being processed locally or sent to a server. The microphone icon looks the same in both cases. You have no way of knowing, in real time, where your words are going.
Feature limitations. Some advanced dictation features - such as dictation in certain app contexts or with specific Siri integrations - may still route through Apple's servers even on supported hardware. The boundaries of what triggers cloud processing are not documented in a way that allows users to predict them.
Processes locally when conditions are met, but falls back to cloud without notice. No visual indicator. Metadata still collected. Improvement programme sends audio samples if opted in. Language and hardware restrictions apply.
All processing happens on-device, every time, for every language supported. No server fallback. No internet connection required. No metadata transmitted. No ambiguity about where your data goes - it stays on your machine.
How to Check if Your Apple Dictation Data Is Being Sent
Before you change anything, it is worth understanding your current configuration. Here is how to audit your Apple Dictation privacy settings.
On macOS (Sequoia / Tahoe)
- Open System Settings → Privacy & Security → Analytics & Improvements
- Check whether Improve Siri & Dictation is enabled. If it is, Apple is collecting audio samples from your dictation sessions.
- Navigate to System Settings → Keyboard → Dictation
- Note which language is set. Cross-reference with Apple's support documentation on on-device supported languages to confirm whether your dictation runs locally.
- Check System Settings → Privacy & Security → Analytics & Improvements → Analytics Data to see what diagnostic data your Mac has been sending.
On iPhone / iPad
- Open Settings → Privacy & Security → Analytics & Improvements
- Toggle off Improve Siri & Dictation
- Open Settings → General → Keyboard → Dictation to verify your language configuration
- If you want to see what Apple has already collected, visit privacy.apple.com and request a copy of your data
What to Look For
When you request your data from Apple, look for entries categorised under "Siri" or "Dictation." These may include anonymised audio references, usage timestamps, and device identifiers. The volume of data can be surprising even for users who thought they were using a "private" service.
How to Stop Apple Dictation From Sending Your Data
If you have decided that any cloud transmission of your voice data is unacceptable, here are your options - ranked from least to most effective.
Option 1: Disable the Improvement Programme
Turn off "Improve Siri & Dictation" in your privacy settings. This stops Apple from collecting audio samples for review. It does not prevent server-side transcription processing if your hardware or language triggers it.
Privacy improvement: Moderate. You stop contributing audio samples, but dictation may still use the cloud.
Option 2: Ensure On-Device Processing
Confirm that your device supports on-device dictation (Apple Silicon Mac, iPhone with A12 Bionic or later) and that your language is in the supported list. Disable Siri to reduce the chance of server-side routing.
Privacy improvement: Significant, but not complete. Metadata and analytics may still be transmitted.
Option 3: Disable Apple Dictation Entirely
Turn off Dictation in System Settings → Keyboard. This is the only way to guarantee that Apple's dictation system is not processing your voice.
Privacy improvement: Complete - but you lose dictation entirely. Unless you replace it with something else.
Option 4: Switch to Fully Offline Dictation
Replace Apple Dictation with a dictation application that processes everything on-device by design - not as a conditional feature, but as its core architecture. This is the approach that gives you both privacy and functionality.
Yaps is built on this principle. Every transcription runs on your device, processed by models that never phone home. There is no server fallback, no improvement programme, no metadata collection, and no internet connection required. It works across all your applications on macOS, with Windows and Android support in development. Your voice data stays exactly where it belongs - on your machine.
If you are concerned about Apple Dictation data but still want to use Siri for non-dictation tasks, you can disable Dictation specifically while keeping Siri active. Open System Settings → Keyboard → Dictation and toggle it off. Then use a dedicated offline dictation application like Yaps for all your speech-to-text needs.
Why "On-Device" and "Offline" Are Not the Same Thing
This distinction matters more than most people realise, and Apple's marketing does not help clarify it.
On-device means the processing model runs locally. It does not mean the application is offline. An on-device model can still send data to the cloud for analytics, quality improvement, error reporting, feature flagging, or A/B testing. The model is local; the application may not be.
Offline means no network communication occurs. The application does not connect to the internet at all during operation. There is no server to send data to, no endpoint to phone home to, no cloud fallback to trigger. Offline is a stronger guarantee than on-device.
Apple Dictation is "on-device" in some configurations. It is never "offline." The application is part of macOS and iOS, which are constantly connected operating systems. There is no way to isolate Apple Dictation's network behaviour from the rest of the system.
Yaps is built differently. It is a standalone application built with Tauri v2 and Rust - not a component of an operating system with its own data practices. It runs on Apple Silicon's Neural Engine for transcription and does not require or initiate any network connections. The offline dictation model is not a feature toggle. It is the architecture.
For users in regulated industries - healthcare, legal, finance - this distinction has compliance implications. A system that "usually" processes locally but "sometimes" sends data to the cloud does not meet the data handling requirements of HIPAA, attorney-client privilege, or financial regulations. A system that is architecturally offline does. If you work in a field where data governance matters, our guide to voice dictation in regulated industries covers this in depth.
What Apple Could Do Better
It is worth noting that Apple's privacy posture is, in many respects, better than its competitors. Google Assistant and Amazon Alexa have historically retained more data, provided fewer controls, and been less transparent about their practices. Apple deserves credit for making the improvement programme opt-in and for investing in on-device processing.
But "better than the competition" is not the same as "good enough." Here is what a truly privacy-respecting dictation system would look like:
- Clear, real-time indicators showing whether audio is being processed locally or sent to the cloud
- Per-session consent rather than a global toggle buried in settings
- Full offline mode that users can enable with confidence, knowing zero data leaves the device
- Open-source models that can be independently verified
- Zero metadata collection during voice interactions
- Transparent retention schedules with specific timeframes, not "shortly after"
Until Apple offers these guarantees, users who care about voice data privacy are left making assumptions about what their devices are doing with their words. And assumptions are not a privacy strategy.
The Bigger Picture: Why Voice Data Is Different
Voice data occupies a unique position in the privacy landscape. Unlike typed text, which is abstract and impersonal, your voice carries biometric information that is inherently tied to your identity. Your speech patterns, accent, tone, cadence, and vocabulary form a signature that is as unique as a fingerprint - and considerably harder to change.
When you type a password and it is compromised, you change the password. When your voice data is compromised, you cannot change your voice. This asymmetry makes voice data fundamentally more sensitive than most other categories of personal information.
This is why the question of whether Apple Dictation sends data to servers matters beyond the immediate privacy concern. It is not just about whether Apple misuses your data - it is about whether that data exists in a location where it could be accessed by employees, contractors, hackers, or government subpoenas. Data that never leaves your device cannot be subpoenaed from a server. Data that does leave your device, even briefly, exists in a context you do not control.
If you are exploring how voice data privacy works more broadly, our analysis of why your voice data is more sensitive than you think goes deeper into the technical and legal dimensions.
Final Thoughts
Apple Dictation is not a privacy disaster. It is a privacy compromise - one that is well-intentioned but structurally limited by its dual nature as both an on-device and cloud-connected service. For casual users who are not handling sensitive information, Apple's current setup may be acceptable. For anyone who considers their spoken words private by default, the ambiguity is the problem.
The good news is that you do not have to choose between privacy and functionality. Fully offline dictation exists today, runs natively on your Mac with under 200MB of RAM, starts in under a second, and works across every application on your system. You hold a key, you speak, and your words appear - without ever leaving your machine.
Your voice is yours. Where it goes should be your choice, not a configuration detail buried three menus deep in System Settings. If you are ready to make that choice explicit, Yaps is here when you are.
Frequently Asked Questions
Does Apple Dictation send my voice to the cloud?
It depends on your configuration. On supported hardware running recent versions of macOS or iOS, basic dictation can process on-device. However, certain languages, older hardware, and Siri integrations may trigger server-side processing. If you have enabled "Improve Siri & Dictation," audio samples are sent to Apple regardless of local processing capability. There is no visual indicator telling you which mode is active during a given session.
How do I turn off Apple Dictation data sharing?
Open System Settings → Privacy & Security → Analytics & Improvements, and disable "Improve Siri & Dictation." This stops audio sample collection. To prevent any server-side processing, confirm your device supports on-device dictation for your language, or disable Apple Dictation entirely and use a fully offline alternative.
How long does Apple keep dictation recordings?
For users who have opted into the improvement programme, Apple retains audio samples for up to six months. A de-identified copy may be retained for up to two years for machine learning purposes. For standard server-side dictation (without the improvement toggle), Apple states audio is deleted "shortly after" processing, though the exact timeframe is not publicly specified.
Is Apple Dictation HIPAA compliant?
Apple does not market Dictation as a HIPAA-compliant service, and its conditional cloud processing makes it unsuitable for handling protected health information (PHI) without additional safeguards. Healthcare providers who need dictation should use a solution that is architecturally offline, ensuring no patient data is ever transmitted. Our HIPAA-compliant dictation guide covers this in detail.
What is the difference between on-device and offline dictation?
On-device means the speech recognition model runs locally on your hardware. Offline means the entire application operates without any network connection. Apple Dictation is on-device in some configurations but never fully offline - it is part of a connected operating system and may still transmit metadata, analytics, or fall back to cloud processing. A fully offline dictation application like Yaps does not initiate any network connections during operation.
Can Apple employees listen to my dictation?
After the 2019 controversy, Apple changed its policy so that human review of audio samples only occurs if you have explicitly opted in through the "Improve Siri & Dictation" setting. If you have not opted in, Apple states that no human listens to your recordings. However, automated systems still process server-routed audio, and there is no independent audit verifying these claims.
Does Siri and Dictation use the same data pipeline?
Siri and Dictation share underlying speech recognition technology and privacy settings, but they are not identical in their data handling. Siri requests are more likely to involve server-side processing because they often require internet-connected features (web search, smart home control, etc.). Dictation has a more limited scope and is more likely to process on-device when conditions are met. The "Improve Siri & Dictation" toggle controls audio sampling for both.
What happens to my dictation data if I sell or trade in my Mac?
If you erase your Mac using the standard macOS reset process, locally stored dictation data - including any speech recognition personalisation - is deleted along with the rest of your data. However, any audio samples or metadata previously transmitted to Apple's servers remain subject to Apple's retention policies and are not affected by wiping your local device. To address server-side data, visit privacy.apple.com and request deletion before transferring your machine.
Are there dictation apps that never send data to the cloud?
Yes. Fully offline dictation applications process all speech recognition locally and do not require or use an internet connection. Yaps is one such application - built with Tauri v2 and Rust, it runs entirely on your device using Apple Silicon's Neural Engine on macOS. No audio, transcription text, or metadata is transmitted anywhere. You can read more about how local Whisper-based transcription works on our blog.
How do I know if my dictation is being processed on-device or in the cloud?
Unfortunately, Apple does not provide a real-time indicator. The dictation interface looks identical regardless of whether processing is local or server-based. Your best approach is to verify that your hardware supports on-device processing (Apple Silicon Mac or iPhone with A12 Bionic or later), confirm your dictation language is in Apple's supported list, and disable "Improve Siri & Dictation." Even then, you cannot be certain that every dictation session is fully local.