Google's Gemini 3.5 Transcribe brings faster, polished speech-to-text to more devices

Google is introducing Gemini 3.5 Transcribe, a fresh voice-to-text model that goes beyond simple word recognition to transform raw speech into polished, edited text. The technology already powers the Rambler voice feature on the Pixel 11, and the company plans to roll it out across its broader product ecosystem.

Performance gains over existing systems#

The new model significantly outpaces Google's previous transcription engine, Chirp 3. Gemini 3.5 Transcribe delivers roughly 70 percent faster processing from voice input to completed text, and its live-speech error rate has improved to 5.5 percent—a meaningful step forward from Chirp 3's 7.32 percent. While the accuracy jump may seem modest, reducing transcription mistakes makes voice input more practical for everyday users.

Smart editing and language support#

What sets Gemini 3.5 Transcribe apart is its ability to refine speech in real time. The model can strip away verbal filler like "ums" and "uhs" that naturally pepper casual speaking, and it handles on-the-fly corrections as users catch and fix themselves mid-thought. The system also taps into custom vocabulary lists to navigate specialized terminology and industry jargon. Support extends across 85 languages and handles up to three concurrent speakers in pre-recorded audio.

Trade-offs in automation#

The convenience of automated cleanup comes with a caveat: because the AI actively rewrites what users say rather than transcribing it verbatim, the final output may drift from the original wording. For brief voice notes and casual dictation, this refinement typically feels natural and helpful, but in contexts where precise word-for-word reproduction matters—legal recordings, interviews, or formal documentation—the intervention could pose problems.