Google unveils new Gemini model that turns rambling speech into structured text
Google has introduced Gemini 3.5 Transcribe, a new audio model that converts unstructured speech into formatted text and will soon work in any web field in Chrome. The model recognizes more than 85 languages and can distinguish up to three speakers.
Google has announced Gemini 3.5 Transcribe, a new AI audio model joining the Gemini Audio family alongside Gemini 3.5 Live and Gemini 3.5 Live Experimental. The company says the new model delivers more accurate speech recognition than earlier versions and can automatically detect over 85 languages.
One of the model's key features is its ability to turn a user's unstructured, natural speech into clean, formatted text. It can handle self-corrections, strip out filler words, and preserve the speaker's intended meaning. Users will also be able to make voice-command edits to transcribed text.
Additional capabilities
Google says Gemini 3.5 Transcribe can learn a user's custom vocabulary and unique spellings, and is particularly good at accurately capturing alphanumeric strings such as order numbers or postal codes. The model can also attribute speech to up to three different speakers with word-level timestamps when working from pre-recorded audio, which could be useful for tasks like transcribing a podcast.
The new model already powers the Rambler feature on Android devices, including Pixel 11-series phones, as well as the Gemini app on macOS, where it can work alongside other Gemini models to carry out more complex, agentic tasks.
Google also said that speech-to-text will soon be usable in any field on a web page through Chrome, allowing users to dictate replies and posts or issue voice prompts to Gemini. The model is already available on Google Antigravity, the company's agentic development platform, and is coming soon to Search Live, Gemini Live, Docs, Keep and Gmail. Developers will be able to access the model through APIs.
