Google is shipping a new voice model while everyone’s still refreshing for Gemini 3.5 Pro. The company announced Gemini 3.5 Transcribe, a speech-to-text system that doesn’t just hear what you say—it edits it. The model strips out filler words, fixes mid-sentence corrections, and outputs what Google calls polished AI text. If you’ve used the Gboard ‘Rambler’ feature on a Pixel 11, you’ve already touched this tech. But it’s now rolling out across Google’s broader ecosystem.
The performance numbers are concrete. Google says Transcribe is about 70 percent faster from speech to final text than its previous engine, Chirp 3. The live-speech error rate dropped to 5.5 percent, compared to Chirp 3’s 7.32 percent. That’s a modest accuracy gain—less than two percentage points—but anyone who’s dictated a paragraph and then tapped through a minefield of typos knows the real win is speed. Fixing transcription errors on a phone is tedious. Getting to usable text 70 percent faster changes how people actually use voice input.
Transcribe also pulls from your custom vocabulary to handle specialized jargon, supports 85 languages, and can process up to three speakers in pre-recorded audio. That speaker separation matters for meeting notes and interviews, where Chirp 3’s limitations were most visible. The model’s ability to restructure your sentences on the fly—removing ‘ums,’ resolving self-corrections—is genuinely useful for dictation, but it’s also a philosophical shift. The AI isn’t transcribing anymore. It’s interpreting.
That’s the tradeoff, and it’s one worth being honest about. The model does technically change your wording, and for short blocks of text Google’s own testing suggests it handles verbal stumbles well. But in legal, medical, or journalistic contexts, an AI that silently rewrites what you said is a liability, not a feature. Google hasn’t said how users will know when text has been edited versus transcribed verbatim—and until that’s clear, this is a tool for casual dictation, not the record.
