Google has updated its audio tools with Gemini 3.5 Transcribe, a new transcription model built to clean up spoken audio while keeping it useful for editing and review. The model can detect specialized jargon, work with more than 85 languages, and remove filler words such as “um” and “uh.” For users who want transcripts that read more like finished text than rough notes, Gemini 3.5 Transcribe is a notable step forward.

The update matters because transcription is no longer just about turning speech into text. Google says the model can automatically format text, support voice-based editing, and adapt to custom vocabulary so unique spellings and technical terms do not need to be corrected by hand. That makes Gemini 3.5 Transcribe more practical for meetings, interviews, lectures, and other audio where accuracy and readability both matter.

Gemini 3.5 Transcribe

Google says Gemini 3.5 Transcribe is an advance over its previous transcription model, Chirp 3, with improvements in multilingual performance and wording error rates. In practice, that means the system is meant to do better with mixed-language audio and with terms that often trip up general-purpose speech recognition.

One of the more useful additions is the ability to supply a customized vocabulary. Instead of forcing users to manually fix names, product terms, acronyms, or domain-specific language after the transcript is generated, the model can adapt to those spellings up front. That could save time for teams working in medicine, law, engineering, research, or other fields where jargon is common.

The model also supports speech attribution for up to three speakers in pre-recorded audio and provides word-level timestamps. Those features are important for anyone trying to review a conversation, follow who said what, or locate specific parts of an audio file without scrubbing through the entire recording.

What Google Says Changes For Users

Google describes the model as one that lets users “edit naturally with just your voice,” and its cleanup features suggest a workflow focused on fast drafting rather than exact preservation of every spoken fragment. By automatically removing filler words and formatting the text, the tool is aimed at producing transcripts that are easier to read immediately.

That can be especially helpful for journalists, content teams, and workers who use audio notes as a first draft. It also raises an important practical point: a transcript that reads better is not always the same as a verbatim record. Users will likely need to decide when they want a polished draft and when they need a word-for-word record of what was said.

Google also mentioned Gemini 3.5 Live and Gemini 3.5 Live Experimental alongside the transcription update, though the company later clarified that those models are not launching yet and did not give a new release date. Based on Google’s earlier information, Live is designed to improve mid-sentence interruptions, language recognition, and live visual processing, while Live Experimental is supposed to narrate its reasoning step by step on more complex tasks.

Where It Is Available Now

Gemini 3.5 Transcribe is rolling out in English starting today for all macOS Gemini app users and for the Rambler dictation feature on Android in select countries and languages. It is also available to developers in public preview through the Gemini API in AI Studio and Antigravity.

Google says Chrome support is coming soon, which could broaden access further once it arrives. For now, the rollout suggests Google is testing the model across consumer apps and developer tools at the same time, giving it a chance to prove itself in both everyday dictation and app-building workflows.

What readers should watch next is whether Google expands language support, how quickly Chrome integration arrives, and whether the company follows through with the delayed Gemini 3.5 Pro launch. For now, the transcription update is the clearest Gemini 3.5 release to reach users, and it points to Google’s broader effort to make audio tools more practical, more editable, and less manual.