Does the Web Speech API send audio to Google?

By default in Chrome, the Web Speech API sends the user's audio to a server for recognition. MDN says so directly ¹. Google's role is less direct: a Mozilla wiki page, written by a competitor, compares its own requests to Google with speech input from Chrome ². We did not find a Google document that names the service or says how long the audio is kept. So the answer is yes by default, but Google's role rests on third-party accounts.

The question matters because a website that uses speech recognition may send a visitor's voice to a company the site never mentions. The guide to the Web Speech API covers the interface, and the overview of dictation in apps covers the other routes a user can take.

MDN says the audio leaves the device in Chrome

MDN's note on the interface says that on some browsers, including Chrome, speech recognition on a web page uses a server-based engine, and that the audio is sent to a web service for recognition ¹. The general statement is broader: by default, speech recognition on a web page sends the user's audio to a server-based service ³.

MDN does not name the service in the Chrome note, so the question of who runs it needs other sources.

A Mozilla wiki says Chrome's speech input can be saved

Mozilla's wiki says it has opted out of letting Google store its voice requests. It contrasts this with speech input from Chrome, which can be saved and attached to a user's profile ².

That is a competitor describing how its own design differs from Chrome, so read it as one vendor's account. It does not quote Google's terms for Chrome.

We could not confirm Google's own description

We looked for Google's own description of the speech service and found none in the Chrome privacy whitepaper we read, which has no passage on speech recognition. That means the link to Google rests on third-party accounts, chiefly Mozilla's, because MDN does not name the service. Until Google publishes a retention and data-use statement for this service, do not tell users more than the evidence supports. For instance, do not say that audio is never stored.

On-device mode is a request the page can make

The Web Speech API now has a setting, processLocally. When a page sets it to true, the specification requires recognition to run on the user's device ⁴. When it is false, which is the default, the browser may choose between local and remote processing ⁵. So the default leaves the choice to the browser.

Chrome 139 added on-device speech recognition to the Web Speech API, so a website can make sure that neither audio nor transcribed speech is sent to a third-party service ⁶. MDN still marks the property as experimental and says to check the compatibility table before relying on it in production ⁷.

It depends: which feature, which browser and which setting

• The setting. With processLocally left at its default of false, Chrome may send audio to a server. With it set to true, the specification requires recognition on the device, so the page should handle a browser that cannot do that.
• The browser. We found the on-device mode in Chrome's release notes, and no matching note for other browsers. Test each browser you support rather than assuming the same path.
• A different Chrome feature. Live Caption in Chrome is a separate feature. In a 2021 announcement, Google said its captions are generated on the device, so the audio does not have to leave the computer ⁸. That statement does not describe the Web Speech API, so do not use it as evidence for a page's speech recognition.

What to do

1. Assume audio leaves the device in Chrome unless you set processLocally. Tell visitors that recognition may use a server.
2. Set processLocally to true for sensitive audio. Medical, legal and personal dictation are the cases where this matters most.
3. Handle the refusal. If the browser cannot run on the device, warn the visitor rather than falling back silently to a server.
4. Check the actual path in each browser you support. Record which path each browser took in your testing.
5. Explain it in your privacy policy. Say whether speech recognition runs on the device or on a server, based on what the browser does, not on what you hoped it would do.

For what the built-in Chrome options do offline, see whether Chrome can do speech to text offline. For the browsers that support the API, see which browsers support speech recognition. The glossary entry on on-device speech recognition explains the term used above.