Google just shipped a suite of Gemini Audio models — 3.5 Live, 3.5 Live Experimental, and a brand-new 3.5 Transcribe — that push voice AI past the point where background noise and awkward pauses derail a conversation. The headline feature is automatic filler-word removal. Say “um” or “uh” mid-sentence and the transcription quietly edits it out. That’s a small thing that makes a massive difference for anyone who’s ever read an AI transcript of a real meeting and cringed.
3.5 Transcribe succeeds Chirp 3, and Google claims it’s a “major advancement” on multilingual performance and word error rates. It detects over 85 languages and can adapt to specialized jargon through a customized vocabulary — so medical terms, product names, or internal acronyms don’t get mangled into nonsense. It also attributes speech for up to three speakers in pre-recorded audio and provides word-level timestamps. For podcasters, journalists, and support teams, that’s a meaningful upgrade from the single-speaker transcriptions most tools still default to.
The two Live models handle real-time voice interaction. 3.5 Live improves mid-sentence interruption handling, language recognition, and live visual processing, while the Experimental version goes further by narrating its reasoning step-by-step as it tackles complex tasks. That last part is genuinely interesting — it turns the model’s internal processing into something observable, which could build trust or just add noise depending on the use case.
Rollout starts today for macOS Gemini app users in English, with Android’s Rambler dictation following in select countries. Developers get public preview access through the Gemini API in AI Studio and Antigravity, and Chrome support is coming soon. I’m most curious whether the jargon customization actually holds up in practice — that’s the kind of feature that sounds great in a demo but can fall apart in the messy reality of accented speech, overlapping voices, and domain-specific shorthand. Google’s timing is also notable: these audio models are shipping while we’re still waiting for Gemini 3.5 Pro, which Google promised back in June. The company is clearly prioritizing voice as the front door to its AI, even as its flagship text model remains overdue.
