Google rolls out Gemini Audio upgrades that erase filler words and boost multilingual transcription

Google announced a set of Gemini 3.5 Audio enhancements, adding Live, Live Experimental, and the first-ever Transcribe model to its portfolio. The Transcribe variant claims a substantial leap over the prior Chirp 3 system, especially in handling multilingual input and reducing wording errors. It automatically excises filler utterances such as “um” and “uh,” formats spoken text, and supports user-defined vocabularies to preserve specialized terminology.

The model can also attribute speech to up to three speakers in pre-recorded audio and supplies precise timestamps for each word. Gemini 3.5 Live improves handling of mid-sentence interruptions and visual processing, while the Experimental version narrates its reasoning steps in real time. The rollout begins today for English users on macOS and for Android dictation in selected regions, with API access offered in public preview and Chrome support slated for the near future.