Google says its latest Gemini transcription model can turn your ramblings into structured text
Developed over time first seen 2 months ago
Google has announced Gemini 3.5 Transcribe, an AI audio model intended to improve speech recognition and turn informal, self-correcting dictation into clean, structured text. The company says it could make voice input more useful for writing, replying and prompting AI tools, although it has removed some previously included information that was not publicly announced.
Google says the model can automatically detect more than 85 languages, remove filler words, learn custom vocabulary and recognise alphanumeric details such as order numbers and postcodes. It can also identify up to three speakers in pre-recorded audio with word-level timestamps, and is already used in Android’s Rambler feature and the Gemini macOS app; planned integrations include Chrome, Search Live, Gemini Live, Docs, Keep, Gmail and developer APIs.
- Google unveiled a model for cleaner, more structured voice transcription.
- It supports over 85 languages and custom vocabulary.
- Google removed some unannounced details from its original announcement.
New here? Start with this
Speech-to-text software converts spoken words into written text. It is used in phone dictation, voice notes, captions and tools that help people write without typing, but it can struggle when speakers pause, change their mind, use unusual names or switch languages.
Google develops Gemini, a family of artificial intelligence models used across its products. Transcription models are designed to recognise speech and, in some cases, organise it by removing verbal fillers, adding punctuation and separating different speakers.
The technology matters because it could make voice input more useful in everyday apps and workplace tools. It also raises practical questions about accuracy, the handling of recorded speech and whether users can check or correct what the software has written.
Full account
Google has introduced Gemini 3.5 Transcribe, an artificial-intelligence model intended to make spoken input more useful by converting it into cleaned-up, structured text. Rather than merely producing a verbatim record, the company says it can accommodate false starts and self-corrections, remove verbal fillers and format the result in line with a speaker’s apparent intent and style.
Google says the model can recognise more than 85 languages, learn user-supplied vocabulary and unusual spellings, and handle details such as postcodes and order numbers. For recordings, it is also said to distinguish up to three speakers and attach timestamps at word level. These features could be useful for tasks ranging from dictated messages to transcribing interviews or podcasts.
The technology already underpins Rambler, the Gboard voice-input feature on Pixel 11 phones, and is being added to the Gemini app for macOS. Google also says developers can use it through the Gemini API and in its Antigravity development platform. A wider rollout is planned for products including Chrome, where the company says users will eventually be able to dictate into web-page text fields, as well as other Google services.
The promise of polished output also raises a limitation: the system may alter what a person literally said in order to improve readability. That could be helpful for informal drafting, but may be unsuitable where an exact transcript matters. Google’s claims on speed, recognition quality and language support have not, in the supplied reports, been independently verified.
Where outlets differ
Source 1 concentrates on the breadth of Google’s planned product integrations and its claims about vocabulary, alphanumeric strings and speaker attribution. It also contains a likely wording error, referring to text-to-speech in Chrome while describing voice dictation.
Source 2 supplies comparative performance figures, saying Google reports a 5.5% live-speech error rate versus 7.32% for Chirp 3 and roughly 70% faster output. It also includes the writer’s limited experience with Rambler and more directly questions whether AI-polished wording is appropriate in every setting.
Both reports say Chrome support is coming soon and describe Rambler as already available on Pixel 11 devices, but Source 2 gives more detail on developer tools and says Antigravity access can use screen context and chat history with permission.