Is Voice to Text Safe? How to Dictate Privately

Voice to text is safe when your words are recognized on your own device and the app keeps neither the audio nor the transcript. Many dictation tools do not work that way by default. Some send each recording to a server, some keep audio or text for months, some use it to improve AI models, and some have people review samples. You can check all four in about five minutes, using the test and settings below.

Voice to text is as private as the place your audio is processed

Your dictation is only as private as the place where recognition happens. Recognition that runs on your device, known as on-device speech recognition, keeps the audio on your hardware. Recognition that runs on a server sends the audio across the network to another company's computers, and that company's rules then apply to it.

Apple's Mac documentation is a clear example of the on-device case. Apple says that on a Mac, voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers ¹. Keyboard settings also show whether dictated audio and transcripts are processed on the device or sent to Apple's servers ².

The server case is common on phones and in browsers. Google's App Store listing for Gboard on iPhone says that voice input from the Gboard microphone is sent to Google for transcription ³. Chrome 139 added on-device speech recognition to the Web Speech API, so a website can make sure that neither the audio nor the transcribed speech goes to a third-party service ⁴.

Windows splits the two modes. Microsoft says device-based speech recognition in Windows 11 processes your voice locally and sends no voice data to Microsoft ⁵. Online speech recognition sends voice data to Microsoft only to provide the service and create transcriptions ⁶. Voice typing uses the online mode ⁷, and when online speech recognition is off, only device-based features such as Narrator are available ⁸. On Windows, then, the private setting can cost you voice typing, which is a trade-off you should choose on purpose.

Table 1. Where recognition runs in common products, as of October 2026

Table 1 shows that the same feature can run in either place. Read the setting or the policy for the exact product you use, because a dictation feature inside a larger app inherits that app's rules. The question on whether voice typing saves your audio covers the storage side in more detail.

Your words can go to six places, and each one is a separate decision

A dictation app can send your words to six places: the recognizer, the network it travels over, the vendor's storage, the vendor's training pipeline, human reviewers and any other company it shares them with. Each place has its own setting, and a yes or no to one does not answer the others. A product can run recognition on your device and still keep your history, or send audio to a server and delete it after recognition.

Recognizer. This is the place where your audio becomes text, on your device or on a server. Section 1 covers it.

Network. When recognition runs on a server, the audio crosses the network. In ChatGPT, pressing the microphone icon sends the recorded audio to OpenAI's models to be transcribed, and the returned text can be edited before you send it ¹¹.

Storage. Storage is the question most people mean when they ask whether dictation records them. Apple says dictation sent to its servers is not stored unless you opt in to Improve Siri and Dictation ¹². ChatGPT is different: audio from dictation is kept for as long as the chat stays in your chat history, and deleting the chat removes the audio clip within 30 days, with exceptions for security or legal reasons ¹³. Microsoft stores Copilot conversation activity for 18 months by default ¹⁴.

Training. Microsoft says it uses data from Copilot, including voice and conversation activity, for AI training, except for certain categories of users or users who have opted out ¹⁵. Anthropic says that in its mobile apps it deletes the audio after converting speech to text, and does not keep a copy of the speech or use it for training ¹⁶. The guide to whether dictation apps train AI on your voice explains how to switch training off where it is offered.

Human reviewers. For people who opt in to Improve Siri and Dictation, Apple says only its own employees, under confidentiality obligations, can access the audio ¹⁷. After reports that people were listening to Siri recordings, Apple said in 2019 that it had suspended human grading of Siri requests ¹⁸. Google said in 2019 that language experts review only around 0.2 percent of all audio snippets ¹⁹, and that one reviewer had violated its data security policies by leaking confidential Dutch audio ²⁰. These are 2019 statements, so check each company's current policy. The guide to human review of voice recordings covers what has changed since.

Sharing. Some AI dictation apps keep transcripts to improve their product. Aqua Voice says that for users with Privacy Mode turned off, it may store transcript data on its servers to improve the product ²¹.

On-device recognition removes the biggest risk, but not every risk

On-device recognition removes the server copy of your audio, which is the largest single risk. Three risks remain, and they are the ones the checklist in section 8 is built around.

The first is the first run. Apple says transcription with its SpeechAnalyzer API is entirely on the device, although the language models must first be downloaded ²². A language may therefore need the network once. Some systems also fall back to the network. ChromeOS Dictation runs on a local component, but when that component is not available, it falls back to network speech recognition ²³. The airplane test in section 4 catches both cases.

The second is what happens to the text after recognition. Dictated text goes into whatever app you are typing in, and that app's sync, history and AI features then apply. A third-party keyboard is a separate exposure. Apple's 2014 warning about third-party keyboards, as quoted by TechCrunch, says that if you enable Full Access for a third-party keyboard, its developers are permitted to access, collect and transmit the data you type ²⁴.

The third is what your voice reveals. Speech can potentially identify speakers, which is why the authors of a 2024 study say that sharing voice recordings raises privacy concerns ²⁵. A voiceprint is the pattern a system can match against a person. The guide to whether your voice is biometric data explains when that matters legally.

Text that a model learns from can also come back out. In a 2021 study, an attack on the GPT-2 language model recovered hundreds of verbatim text sequences from its training data ²⁶. That study tested one older model, not dictation tools, so it does not show that your dictation is kept in a model. It does show why a policy that says "we do not train on your data" is worth reading closely, and why the question on whether AI dictation apps send your text to the cloud matters for any tool that adds an AI step after recognition.

Apple's design reasoning is the same as the one here. Apple says on-device machine learning benefits privacy because the data for inference workloads stays on the device rather than on a server ²⁷.

The airplane-mode test shows in thirty seconds where recognition happens

The quickest check needs no settings and no documents. Disconnect from the internet and dictate. If the words appear, recognition ran on your device for that language. Work through these steps:

1. Turn on Airplane Mode on your phone, or disconnect your Mac from Wi-Fi and any network cable.
2. Open the app you use and dictate one sentence in your usual language.
3. If the words appear, recognition ran on the device. If nothing appears, or the words appear only after you reconnect, the app sends audio to a server.
4. Reconnect and check whether the app uploads the text it dictated offline.

The test depends on language. Apple says dictation on iPad is processed on the device in many languages and needs no internet connection ⁹, but dictation may not be available in all languages or in all countries or regions ²⁸. Run the test in each language you dictate in.

A pass tells you that recognition works offline. It does not tell you that nothing is kept afterward. Section 5 covers storage, which the airplane test cannot see. The test works on any dictation app, including Yap, which we make, so run it on ours before you trust it.

A privacy policy must answer four questions: storage, training, human review and sharing

A dictation app's privacy policy is useful only if it answers four questions. Most policies answer them under headings such as "data retention" and "how we use your data", so search for those words first.

Table 2. The four questions a dictation policy must answer

Retention periods differ by product and by account type, and the difference can be large. Microsoft's own figures show it. For work or school accounts, Copilot Voice audio is temporarily stored for feedback and deleted after 48 hours ²⁹, which is two days ³⁰. The default retention for Copilot conversation activity is about 548 days ³¹.

A label is a summary, not a policy. Apple's App Store privacy labels say whether data is linked to you, and Apple defines that as data collected in a way that is linked to your identity, such as to your account, your device, or your details (for example, your phone number) ³². Use the label to see what the developer declares, then read the policy to see how the audio is handled. If a vendor offers zero data retention, where the provider keeps nothing after the request, get that commitment in writing. The checklist for a dictation app's privacy policy turns these four questions into a list you can use on any app.

Every major platform has settings that keep more of your voice to yourself

Defaults change between operating system versions, so each setting below names the version it applies to. Confirm the menu on your own screen, because Apple and Microsoft move options between releases.

Table 3. Privacy settings to check, by platform

Two Apple settings matter most. Deleting Siri and Dictation history needs an internet connection on a Mac ⁴⁰, so do it while you are online. The Improve Siri and Dictation setting controls whether Apple stores and reviews a sample of your audio ⁴¹, so leave it off unless you want to help improve Apple's speech recognition. The guide to Mac dictation privacy walks through the full set of Apple settings, and the guide to Windows dictation privacy covers the Windows options.

Smart speakers follow a different pattern. Amazon told Echo owners in March 2025 that their voice recordings would be sent to and processed in the cloud from March 28, and deleted after Alexa processes the requests ⁴². Android voice features vary. Android Central reported in 2024 that Google says voice message transcripts in Google Messages are processed on the device, and that audio messages are not sent to the company ⁴³. In 2019 Google moved Gboard voice typing onto the phone with an 80 MB speech recognizer that works offline, first on Pixel phones in American English ⁴⁴. On Android, a February 2025 report said Gboard offers offline voice typing on all devices ⁴⁵, so the airplane test in section 4 applies there too. Google and Amazon change their deletion menus often, so the question on deleting voice recordings from Google, Apple, Amazon and Microsoft keeps the current steps in one place.

Patient, client and classified content needs an agreement or an on-device tool

Health, legal and government content adds rules that a consumer policy does not cover. The question changes from "is this app private?" to "who holds this audio, and what did they agree to?"

Health records. HHS says that a covered entity and a cloud service provider that holds its health information must sign a HIPAA-compliant business associate agreement (BAA) ⁴⁶. Encryption does not change that: HHS says that lacking an encryption key does not exempt a cloud service provider from business associate status ⁴⁷. Patient information is protected health information (PHI) when it is held by a covered entity. The guide to HIPAA-compliant dictation covers which products sign the agreement.

Legal work. ABA Formal Opinion 512 concludes that informed consent is required before client information is entered into self-learning generative AI tools whose outputs could disclose it ⁴⁸. Dictating client matters into a cloud AI tool may therefore need the client's consent, not only a privacy setting. The guide to dictation for lawyers explains how attorney-client privilege applies to dictation.

Voice as biometric data. Illinois's Biometric Information Privacy Act defines biometric identifiers to include voiceprints, alongside fingerprints and iris scans ⁴⁹. The GDPR defines biometric data as personal data from specific technical processing of a person's physical, physiological or behavioural characteristics that allows or confirms the unique identification of that person ⁵⁰. Article 9 prohibits processing special categories of personal data, including biometric data used for the purpose of uniquely identifying a natural person, unless an exception applies ⁵¹. A transcript of what you said is different from a voiceprint, which is the feature a system uses to identify you. The guide to biometric data, GDPR and CCPA sets out the tests in more detail.

Recording a call. Dictating into a call is a separate question. California Penal Code 632 punishes a person who intentionally records a confidential communication without the consent of all parties ⁵². The guide to recording consent laws covers the other states and countries.

For regulated content, there are two safe options. Use a vendor that will sign the agreement the regulation requires, or use an on-device tool that keeps no copy at the vendor. If a vendor will not sign, do not send the content to that vendor's servers, and confirm with the airplane test that the tool works offline.

A private dictation setup takes five minutes

For most people, the free dictation built into the phone or computer is enough once the settings above are right. Follow these steps in order:

1. Run the airplane test on the app you use most, in each language you dictate in (section 4).
2. Turn off server-side history and training where the product offers a switch, using the settings in Table 3.
3. Delete existing history. On a Mac, use the Siri and Dictation history steps for your macOS version while you are online.
4. Answer the four questions in the app's policy: storage, training, human review and sharing (Table 2).
5. For AI dictation apps, check whether the audio is kept with your chat history and whether you can delete it. ChatGPT removes a deleted chat's audio within 30 days, unless OpenAI must keep it for security or legal reasons ¹³.
6. Turn off Full Access for third-party keyboards unless a keyboard needs it ²⁴.

If you are choosing an app rather than configuring one, the guide to choosing a dictation app compares the features that matter, and the guide to offline voice typing on iPhone and Android covers the offline options on phones. The short answer is that voice to text is safe when the test passes, the storage and training switches are off, and the policy answers all four questions in writing. For the whole subject, start with the voice privacy and security overview.