Where your voice data goes after you say it

Speaking to an assistant can involve several separate activities: detecting an activation, processing the request, saving a recording or transcript, and using selected data to improve a service. These activities are not interchangeable. Processing audio does not necessarily mean storing the recording, and not storing audio does not necessarily mean keeping no transcript.

The answer depends on the exact product, device, region and settings. This guide uses Google Assistant and Siri as documented examples. Their policies should not be assumed to apply to Gemini, Alexa or every voice-enabled app.

From sound to an answer

A useful conceptual sequence is activation, audio processing, interpretation, a response or action, and any permitted retention. Some systems use speech-to-text as an intermediate step; others may process audio more directly. Treat this as a way to ask questions about a product, not a universal description of its internal architecture.

Ask separately what leaves the device, what remains in account history, what is used for improvement, and what a connected app receives. A single privacy toggle may cover only one of these.

What is stored: two concrete examples

Google says the Google Assistant voice-and-audio activity setting is off unless you enable it. With it off, Assistant recordings are not saved to your Google Account on Google servers under that setting. Turning it off does not delete previously saved recordings, and other settings can save audio elsewhere. Google also notes that settings are being updated, so labels may differ. See Google’s audio activity guidance.

Apple says Siri settings indicate when speech is processed on the device; otherwise audio is sent to Apple for processing. Its notice says Apple does not store audio unless you opt into Improve Siri and Dictation. Transcripts and related request data are a separate category: Apple may retain them for up to two years, and reviewed subsets may be kept longer. See Siri, Dictation & Privacy.

There is therefore no single retention period for all voice assistants. Check the policy for each data category instead of assuming that everything disappears when a recording is deleted.

On-device and cloud processing

Local processing can reduce the information sent away for a particular task. It does not prove that every feature works offline or that no related information is shared. A cloud-supported request may involve remote processing even when another feature of the same device works locally.

Use the device’s own explanation of the specific feature. Test whether the commands you rely on still work after changing settings. Avoid assuming that local processing is always less accurate, or that cloud processing always performs better.

Who may review data?

Human review, service-provider access and optional improvement programs need separate checks. Apple’s notice, for example, says a subset of Siri transcripts may be reviewed even though audio storage has its own opt-in condition. Its notice also distinguishes requests handled by third-party apps, which have their own terms and privacy policies.

Do not assume that every recording is heard by a person—or that turning off one improvement program eliminates every type of review. Look for explicit statements about audio, transcripts, access restrictions and third-party integrations.

A practical privacy check

First, identify the exact assistant and account in use. Shared speakers, phones and connected apps may have separate controls. Read the descriptions before changing an option.

Second, review optional audio saving and improvement settings. Choose whether you want to participate. Then inspect saved history separately; changing a future collection preference may leave earlier records intact.

Third, review deletion controls carefully. Google’s guidance explains that Web & App Activity auto-deletion covers more than audio. Select only the scope and retention period you intend, and read the confirmation before deleting anything.

Fourth, check microphone mute, manual activation and wake-word options. Detection of a wake word and continuous transmission to a server are different things. Avoid speaking passwords or confidential details around a device you do not control.

Finally, repeat this review after major software or account changes. Keep useful features enabled where you are comfortable with the documented data handling; there is no need to assume that every privacy choice has the same functional trade-off.

Frequently asked questions

Does “audio not saved” mean no data is kept? No. A provider may handle transcripts, request metadata and recordings differently. Read each category’s policy.

Does disabling an improvement option delete my history? Not necessarily. Check the history and deletion controls separately, including any exceptions described by the provider.

Can I choose a 30-day or 60-day retention period? Only if the product offers it. Do not assume a particular interval exists; use the choices shown in your account.

What about sensitive conversations? Use an appropriate private channel and avoid activating an assistant unnecessarily. For workplace or regulated information, follow your organization’s approved tools and policies.

One useful step today

Open the assistant’s privacy controls and answer two questions: is optional audio saving enabled, and what earlier history remains? Then decide whether to adjust those controls. This is more reliable than assuming that all assistants record, retain and train on speech in the same way.