Google has replaced the server-based speech pipeline used by Gboard with an end‑to‑end neural recogniser that runs on Pixel phones. The new system uses an RNN‑transducer model compressed to run locally (roughly 80MB), so dictation is handled on the device rather than being sent up to Google’s servers.
What changed
Until now, Gboard's voice typing sent audio to Google's cloud speech servers for transcription, then returned the text to the phone. The new approach puts an entire neural speech recogniser — trained end to end — inside the phone. That means the phone does the heavy lifting of turning spoken words into text, using a single neural model called an RNN‑transducer, rather than a multi-stage server pipeline.
Technically it’s a significant shift: moving from a server-dependent chain of components to a compact, unified model that runs locally. Google has compressed the model so it can fit and run on a Pixel without needing constant network access.
Why on‑device matters
There are three practical benefits you’ll notice right away.
- Lower latency: because audio no longer has to travel to and from the cloud, transcriptions appear much more quickly. That gives a more natural, near‑instant dictation experience.
- Works offline: dictation will function even when you don’t have a data connection, which is useful on the move, in areas with poor reception, or when you want to avoid using mobile data.
- Privacy: with recognition happening locally, there’s no need to stream what you say to remote servers for processing. For many people that reduces the amount of personal audio and text leaving their device.
Running neural models on the phone also means fewer server resources and less dependency on a network, so the system is more resilient when coverage is patchy.
Availability
Google is rolling the new on‑device recogniser out to Pixel phones first, with US English supported initially. Other languages and devices may follow in time, but the first users to see the change will be those on Pixel hardware in the United States.
If you use Gboard on an eligible Pixel, you should notice faster responses from voice typing and the ability to dictate without a connection. This is part of a broader trend to push machine learning work from the cloud into phones — trading some extra local storage and compute for speed, reliability and privacy.
This article has been restored to the What's New On The Net archive as part of the site's relaunch.
