AI Radar · Daily · August 27, 2026
Google released Gemini 3.5 Transcribe on August 26, 2026, replacing Chirp 3 as its speech-to-text model, and, unusually for a launch post, backed the release with a third-party benchmark rather than only its own claims. According to Google's announcement, Artificial Analysis measured the model at 4.0% word error rate (WER) streaming and 2.6% non-streaming, with time to final transcription 70% faster than Chirp 3. On the FLEURS multilingual benchmark, Google reports 5.50% WER streaming and 5.04% non-streaming.
The model is live today, not on a roadmap: it's in public preview through the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform, with narrower rollouts in the Gemini app for macOS (English only) and the Rambler dictation feature on Android in select countries. Chrome talk-to-type integration and a Gemini Enterprise Customer Experience release are marked "coming soon," meaning they weren't usable at launch.
Two model endpoints ship: gemini-3.5-transcribe for pre-recorded files (up to one hour, or 30 minutes with diarization and timestamps enabled) and gemini-3.5-transcribe-live for real-time streaming with sub-second latency, capped at 10-minute sessions, per Google's API documentation.
What actually changed from Chirp 3
The headline change isn't raw accuracy, it's what the model does with the transcript. Chirp 3 returned speech verbatim, "as heard." Gemini 3.5 Transcribe applies what Google calls "smart transcription": it strips filler words ("um," "ah"), auto-formats the output, and resolves self-corrections in real time, so "let's meet Tuesday, no, Wednesday" collapses to the corrected statement rather than transcribing both halves. It also accepts custom vocabulary lists, up to 1,000 terms per request according to the API docs, so specialized jargon and unusual spellings can be biased toward correctly. Language coverage is broad: 85-plus languages with automatic detection and regional accent handling, versus a narrower set on Chirp 3.
Speaker attribution is where the messaging gets less clean. Google's launch post describes attribution "for up to three speakers," with anything beyond that labeled experimental. The API documentation, however, lists diarization support for up to 8 speakers on the file-processing endpoint, a gap between the marketing framing and the technical ceiling that developers relying on the feature for multi-speaker calls should test before trusting. Diarization and word-level timestamps are also file-only: the live streaming endpoint supports neither.
The catch: independent vs. self-reported numbers
The 4.0%/2.6% WER figures are attributed to Artificial Analysis, an independent benchmark firm, a step up from vendors quoting only in-house numbers. But the FLEURS multilingual figures (5.50%/5.04%) come from Google itself, and Google hasn't published a side-by-side WER table against Chirp 3, only the aggregate "70% faster" latency claim, so the accuracy delta over its own predecessor is never directly stated. Reaction on the Hacker News thread is mixed: commenters said the model "beats every other model on accuracy" but "needs more work on latency" for interactive use, and that it can miss business-specific vocabulary in code-switched speech, an area where Voxtral Mini 3B and Soniox v5 reportedly still edge it out.
Pricing runs roughly $0.005 per minute blended for batch transcription and $0.009 per minute for live streaming, per Google's API documentation, which also confirms the model is API-access only, no open weights, no self-hosting. Caching, function calling, and the Batch API are all explicitly unsupported in this preview.
What it changes for builders
The practical upgrade for anyone shipping transcription today is speed and cleanup, not a step change in accuracy over what streaming ASR already does well. The 70% latency gain matters for live captioning and voice-agent pipelines where time-to-final-text is the bottleneck, and built-in filler-word removal saves a post-processing step many teams currently script by hand. But the preview label matters: production SLAs and Chrome-level browser integration aren't settled, and the speaker-count discrepancy between the blog post and the API docs is worth testing against your own audio before committing a multi-speaker workflow to it.
Sources
- Intelligent transcription with Gemini 3.5 Transcribe - Google Blog — 2026-08-26
- Gemini 3.5 Transcribe - Gemini API Docs — 2026-08-27
- Gemini-3.5-Transcribe - Hacker News discussion — 2026-08-27
- Google AI Releases Gemini 3.5 Transcribe - MarkTechPost — 2026-08-27