Google has named the speech model that has been quietly running Rambler on your phone. Gemini 3.5 Transcribe, announced on 26 August, is the transcription engine behind Rambler in Gboard on Android and behind voice input in the Gemini app on macOS — and Google says it is coming to Chrome, where it will handle dictation into any text field on the web.
What Gemini 3.5 Transcribe changes on Android
Nothing new appears on your phone today. Rambler, the Gboard feature that turns rambling speech into cleaned-up text, already uses this model — Google’s announcement says consumers are already benefiting from it. What the post does is put specifics behind a feature that shipped without them.
On Gboard, Google says 3.5 Transcribe converts spoken thoughts into formatted text with filler words stripped out, and that you can use your voice to make edits, fix misspellings and change the writing style. It also says the model handles self-corrections in speech — Google’s own example is let’s meet Tuesday, no, Wednesday — and cleans up disfluencies rather than transcribing them literally, which is the practical difference between old dictation and this generation of it.
For anyone outside the English-speaking market, the number that matters is language coverage: Google says the model automatically detects and transcribes over 85 languages and handles regional accents and dialects. Accent handling is the single most common complaint about voice typing in India and across Southeast Asia, and it is also the claim Google offers no per-language breakdown for.
Chrome dictation is the next visible change
The Chrome integration is the part that has not shipped. In Google’s words:
“Coming soon to Chrome, you’ll be able to talk to type in any web field — making it effortless to dictate replies, draft posts, or prompt Gemini in Chrome more naturally and easily with your voice.” — Google, announcing Gemini 3.5 Transcribe
Google gives no date, no rollout regions and no indication of whether it is desktop-first or reaches Chrome on Android at the same time. That last question is the one that decides whether this is a phone feature or a laptop feature, and the announcement does not answer it. Chrome on Android has already been the delivery route for Gemini in Chrome, which went live for US Android users earlier, so there is a precedent for the mobile browser getting these features — but a precedent is not a commitment.
The accuracy numbers, and who measured them
Google’s headline figures are worth reading with the attribution attached, because Google itself credits an outside firm rather than its own testing:
- Word Error Rate: an average of 4.0% streaming and 2.6% non-streaming, as measured by Artificial Analysis.
- Multilingual WER: 5.50% streaming and 5.04% non-streaming on the FLEURS benchmark, across what Google describes as a set of top languages and locales.
- Speed: time to final transcription improves by 70% over Chirp 3, Google’s previous transcription model — again as measured by Artificial Analysis. This is a latency figure for one specific metric, not a blanket “70% faster” claim.
- Speakers: speech attribution with word-level timestamps for up to three speakers in pre-recorded audio, with support beyond three marked experimental.
A benchmark run at a vendor’s request on a vendor’s chosen test set is a data point, not an independent verdict. The useful comparison for a phone owner is not 4.0% versus 2.6% but whether Rambler stops mangling proper nouns in your own accent, and no published WER answers that.
The caveats Google’s post leaves out
Two things a reader should hold onto. First, this is a Gemini model exposed through the Gemini API, and Google’s announcement describes cloud APIs throughout and does not say transcription runs on the device — so assume your audio is being processed by Google’s servers unless and until Google says otherwise. Google also notes that on its Antigravity tool the model pairs screen context and chat history with the user’s permission, which is a reminder that these voice surfaces increasingly read more than what you said.
Second, capability announcements are not entitlement announcements. Rambler on Android already has a usage limit Google has declined to put a number on, and nothing in this post changes that. A better model behind the same undisclosed cap is still a cap.
Developers can use the model now through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, via a Live API endpoint for real-time streaming and an Interactions API endpoint for pre-recorded audio.
