EU Orders Google to Open Android Voice Features to Rivals

Say "Hey Google" to an Android phone and it wakes up, having spent all day listening without noticeably touching the battery. Say "Okay Nabu" — the wake phrase for Home Assistant, a popular piece of free software that lets people run their smart lights, locks and thermostats from a computer in their own house rather than a company's cloud — and the same phone would run its battery down roughly fifteen times faster doing the same job, keep a green "microphone in use" dot burning on screen all day, and force the owner to give up Google's own assistant entirely in exchange.

That asymmetry was not an accident of engineering. On 16 July 2026, the European Commission, the EU's executive arm, adopted a decision ordering Alphabet, Google's parent company, to open up eleven Android features to competing apps on what the regulator calls "equally effective" terms — free of charge, with documentation and testing tools, and without any commercial agreement required. Among them: always-on wake word detection, access to ambient data from the microphone and camera, control of on-screen actions, access to on-device AI models, and structured hooks into Google's own apps such as Gmail, Calendar and Maps.

The decision came out of the Digital Markets Act, the 2022 EU law that singles out a handful of very large platforms — it calls them "gatekeepers" — and imposes obligations on them that ordinary companies do not face. Before adopting the decision, the Commission ran a consultation, and one of the parties it invited was the Open Home Foundation, the non-profit behind Home Assistant, whose Android developer had spent three years failing to make a wake word work properly on the platform.

The chip Google reserved for itself

The technical detail matters here, because it explains why the gap was so hard to close from the outside.

Modern phones detect their wake phrase in two stages. The first stage runs on a digital signal processor, or DSP — a small dedicated chip built to chew through audio at a tiny fraction of the power the phone's main processor would need. It runs a deliberately crude model, and it is walled off: the process has no network access and cannot pass any audio onward until it thinks it has heard the phrase. Only then does the main processor wake up and run a bigger, more accurate model to confirm.

Almost every current Android phone has such a chip. Almost no app can use it. Access was restricted to Google and to the handset manufacturer, and the documentation was never published — the Home Assistant developers worked out how the mechanism functioned by reading Android's source code.

So they built the only thing they could: a small detection model running continuously on the main processor inside their own app. It worked, at a price. Battery consumption attributable to the app jumped from around 1% to around 15%. The privacy indicator stayed lit permanently, because the app genuinely did hold full microphone access the whole time — and, unlike Google, it had no way to offer the stronger promise that the audio was locked in a sandbox that physically could not send it anywhere. Users simply had to take the developers' word for it. And to survive a reboot at all, Home Assistant had to be set as the phone's default assistant, which meant giving up Gemini, and with it voice calling, messaging and everything else wired to that role.

Decoupling the default

Two lines in the decision do most of the work. The first says Alphabet may not condition access to these features on an app holding a default role, including the default assistant role. The second requires that wake word detection from multiple services — third parties and Alphabet alike — be able to run at the same time.

Taken together, that would end the either/or. A phone could listen for "Hey Google" and "Okay Nabu" simultaneously, routing each to a different assistant, with neither one demoted. It also opens the door to something home automation enthusiasts have wanted for years: different phrases for different purposes, one for the household, one for the person who administers the system.

Whether the low-power chip can comfortably listen for several phrases at once is a genuine open engineering question. These processors were designed to detect exactly one thing while doing as little work as possible; watching for more means more work on the chip and more false wake-ups handed to the main processor.

Google calls it a security risk

Alphabet has objected publicly, warning that the decision hands third parties sensitive device permissions and risks exposing user data without people's knowledge or consent.

The counter-argument is straightforward. The decision leaves the safeguards intact: Google can still require explicit user consent, still show the privacy indicators, still let people revoke access app by app, and still gate the most sensitive categories — health data, for instance — behind Play Store review for security and data minimisation. The capabilities themselves already exist and are already used by Google's own services, which do not ask permission of anyone. What the ruling changes is not the level of risk on the device but who decides which companies get to take it.

There is a plausible privacy gain, too. Today an independent assistant must ask users to trust it with a live microphone. Under the decision, first-stage listening happens in the same isolated, network-blocked process Google reserves for itself — a guarantee enforced by the operating system rather than by a promise in a privacy policy.

Two years before anything ships

None of this arrives soon. Alphabet must deliver the changes in Android 18, the next major release, by 1 August 2027. Concurrent wake word detection — the multiple-assistants-at-once part — is deferred to Android 19, no later than 1 August 2028.

And the Commission has told Google what to build, not how. That leaves room for what regulators have learned to call malicious compliance: an interface that technically satisfies the text and is miserable to use in practice. The decision anticipates it. "Equally effective" is defined to cover ease of use, speed and energy consumption; documentation and testing tools must be published; and Alphabet has to report its progress to Brussels every month. Shipping the bare minimum is still possible. It is just now something that has to be explained, in writing, thirty days at a time.