Question · Voice privacy and security
Can you use speech to text for confidential documents?
Yes, for many documents, but only with a tool that processes speech on the device or a cloud service your organization has approved for that content. Dictation that stays on the computer avoids the main risk, because the audio never reaches a server. Cloud tools also need a contract, retention controls and a check of their default settings, and privileged or regulated records need more than that.
Updated October 17, 20264 min read8 cited facts, each traced to its source
Yes, but the answer depends on the tool, not on the fact that you are speaking. Speech is data. A dictated client name, deal term or patient note carries the same confidentiality as a typed one. What changes is the route: whether the audio stays on your computer, goes to a server, or is kept or reviewed afterward. Our guide to dictation at work covers how organizations approve tools, and this page covers the document question.
Tools that process speech on the device keep audio off the network
The lowest-risk option is on-device speech recognition, where the engine runs on your own computer. On a Mac, Apple says voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers 1. That statement covers general text, such as messages and notes, and not dictation in a search box.
The rule is not automatic. Apple also says Dictation on a Mac may need an internet connection, and that Keyboard settings show whether your dictation is processed on the device 2. Check that setting on the Mac you will use, rather than assuming it.
On Windows, device-based speech recognition processes the voice locally and sends no voice data to Microsoft 3. Microsoft does not say which third-party apps use that mode, so test an app's network traffic before you rely on it.
Some built-in and cloud tools send speech to a server
Not every built-in tool works this way. Windows voice typing needs an internet connection, a working microphone and the cursor in a text box before it will type what you say 4. Because it needs a connection, it cannot be used offline for confidential work.
Some AI tools send audio to their own servers by design. Claude Code's voice dictation streams the recorded audio to Anthropic's servers for transcription, and the audio is not processed on the computer 5. A cloud dictation service sends audio for recognition by design, so for that kind of tool the question moves to the contract and the settings.
Storage and human review are the risks that remain
Even when a tool sends speech to a server, the risk is what happens next. Samsung's 2023 concern was that data sent to AI platforms is stored on external servers, where it is hard to retrieve and delete 6. That concern is one reason organizations often require a contract and a retention setting before a cloud tool touches work content.
Human review is a separate risk, and it can come from a setting you turn on. Apple says that for people who opt in to Improve Siri and Dictation, only Apple employees, under strict confidentiality obligations, can access the audio 7. Leave that option off for confidential material unless your organization has approved it.
It depends on the document and the rules that apply to it
The tool is half the decision. The material is the other half.
- Privileged communications. Dictating legal advice or a client's matter raises questions about whether the tool's terms keep privilege intact. We cover them in dictation and attorney-client privilege.
- Health records. HHS says that a covered entity and a cloud service provider holding its health information must sign a HIPAA-compliant business associate agreement 8. Read whether on-device dictation is HIPAA compliant before you use any tool for patient notes.
- Government and classified material. Our sources do not cover classified systems. Ask your security office which systems are authorized for each level, and do not assume that a commercial dictation app qualifies.
What to do before you dictate a confidential document
- Classify the document. Decide whether it is public, internal, confidential, privileged or regulated. The label decides which tools you may use.
- Confirm where speech is processed. On a Mac, read the Keyboard settings entry for Dictation. On Windows, check that the speech mode is device-based, and test the app's network use.
- Test with the network off. Disconnect the computer and dictate a passage in each language you need. If the text still appears, recognition is local for that language.
- Turn off review and training. Leave opt-in review settings off unless they are approved, and check any AI training option in the app's settings.
- Get the paperwork for any cloud tool. Ask for a data processing agreement and a list of subprocessors. Our checklist for a dictation app's privacy policy lists the questions to put to a vendor.
- Write the approval down. Record which tool, which kinds of content and which settings are approved. The outline of a dictation policy shows what to include.
Yap, which we make, is an on-device dictation app for Mac and Windows, so it belongs in the first group above. Test it with the steps here before you approve it for confidential work, as you would any other tool.
For the wider picture, read the overview of voice privacy. Related questions: is offline speech recognition more private than cloud, whether your employer can see what you dictate, and whether speech to text can run on an air-gapped computer.
Evidence
Every number on this page, with where it comes from. Open a fact to see the exact words in its source.
- 1.
- 2.
- 3.
- 4.
- 5.
- 6.
- 7.
- 8.
Go deeper
Related questions
- Can government employees use AI dictation?
- Can speech to text run on an air-gapped computer?
- Can your employer see what you dictate?
- What should a company dictation policy include?
- Why do companies ban AI notetakers and dictation apps?
- Are conversations with an AI chatbot privileged?
- Are open-source offline dictation apps like Handy and Murmur safe to use?
- Can a third-party keyboard see what you type?