Gemini 3.5 Transcribe Brings Speech Recognition Into the Gemini Family
Google launched Gemini 3.5 Transcribe — streaming and file speech-to-text with 4.0%/2.6% Artificial Analysis WER, 70% faster time-to-final than Chirp 3, 85+ languages, and public-preview APIs, with use in Gboard Rambler. Google cites AA’s results; noisy call-center audio is the open question.
Google is folding ASR into the Gemini name the way it folded image and video. Google cites Artificial Analysis WER results; whether live holds on noisy call-center audio rather than FLEURS is the open check.
Google announced Gemini 3.5 Transcribe, with its APIs in public preview. Google cites Artificial Analysis WER: 4.0% streaming / 2.6% non-streaming; FLEURS 5.50% / 5.04%. Google: 70% faster time-to-final vs Chirp 3; 85+ languages; up to 3 speakers on files. API IDs: gemini-3.5-transcribe and gemini-3.5-transcribe-live. Already in Gboard Rambler, Gemini macOS, and Antigravity mic.
What shipped#
| Item | blog.google Aug 26 |
|---|---|
| Family | Speech-to-text, not a chat Flash |
| Live + batch | Two API surfaces |
| Languages | 85+ |
| Speakers | Up to 3 on files |
Claims vs checks#
- Artificial Analysis WER (4.0% / 2.6%) is the independent score Google cites alongside FLEURS.
- Google attributes the 70% faster time-to-final comparison with Chirp 3 to Artificial Analysis; its original scorecard was not independently inspected here.
- Whether Chirp disappears from Cloud Speech is not settled by this launch post.
Limits#
- Live WER on noisy call-center audio is not shown here.
- Chirp sunset timeline for Cloud Speech customers is not in the launch note.
Correction and source review — October 14, 2026: The launch announced public-preview APIs, not a Chirp 3 retirement. Qualified performance attribution to Google’s report of AA testing and removed unsupported Pixel 11 specificity.