Gemini 3.5 Transcribe wants you to stop worrying about being a perfect talker

Google's latest speech-to-text model, Gemini 3.5 Transcribe, is designed to handle imperfect speech and convert it into clean, formatted text without expecting flawless delivery. This new model, part of Google's "Gemini Audio" suite alongside Gemini 3.5 Live and Gemini 3.5 Live Experimental, automatically removes filler words such as "um" and "ah," while also accounting for self-corrections in real time.

For instance, if you say "let's meet Tuesday, no wait, Wednesday," the model recognizes your intended time as Wednesday and updates accordingly. You can even dictate edits using your voice. Additionally, Gemini 3.5 Transcribe allows users to generate images or analyze files through function calling, a feature currently available in the Gemini app on macOS.

The model boasts impressive accuracy, with a word error rate of just 4% for real-time streaming and 2.6% for pre-recorded audio, and supports over 85 languages, regional accents, and specialized jargon. It can even label up to three different speakers in a recording with timestamps. Gemini 3.5 Transcribe is already integrated into Rambler, the dictation feature on Android's Gboard, the Gemini app on macOS, and is set to roll out for Chrome, enabling direct dictation into any web field.

Developers can currently access the model in Google AI Studio and Antigravity, Google's agentic coding platform. For enterprise users, it's available through the Gemini Enterprise Agent Platform. The model's capabilities are set to expand further, integrating into Search Live, Gemini Live, Docs, Keep, and Gmail. The release brings Gemini Audio, Google's newest audio-driven AI suite, one step closer to mainstream use.