Google on Wednesday announced Gemini 3.5 Transcribe, an audio model that turns speech into edited text, according to Ars Technica and The Verge. The release arrives while Gemini 3.5 Pro, which Google had pointed to for June, is still outstanding.
Google told The Verge the model is a step up from Chirp 3, its previous transcription engine, especially on multilingual performance and wording error rates. Ars Technica reported Google’s figures: about 70 percent faster from voice to final text, with a live-speech error rate of 5.5 percent, compared with 7.32 percent for Chirp 3.
The model does more than hear words. Google says it can strip filler such as “um” and “uh,” apply a custom vocabulary for specialized jargon, and format text as you speak. Ars Technica said it works in 85 languages and can handle up to three speakers in pre-recorded audio. The Verge added word-level timestamps and speaker attribution on that same three-speaker cap. Users can “edit naturally with just your voice,” Google said, per The Verge.
That cleanup is the product. Ars Technica noted the model technically changes the wording of what you said, which may not fit every use. Short dictation is one job. A verbatim record is another.
Rollout is partial. The Verge said Gemini 3.5 Transcribe is live in English for all macOS Gemini app users and in the Rambler dictation feature on Android in select countries and languages. Ars Technica said Rambler already runs on Pixel 11 Gboard, and that Google will expand it to more Gemini Intelligence devices later this year. Developers get public preview access through the Gemini API, Google AI Studio, and Antigravity. Chrome support is coming “soon,” both outlets reported, which would put voice input in any web field.
The Verge later updated its story. Google had also briefed reporters on Gemini 3.5 Live and 3.5 Live Experimental, then said those models are not launching yet and did not give a new date. Transcribe is the one that shipped.
Google will happily take the ums out of your sentence. Whether 3.5 Pro ever shows up to finish the rest of the 3.5 line is still a separate wait.