Gemini 3.5 Transcribe: 4.0% WER in streaming preview

Google shipped Gemini 3.5 Transcribe on August 26, a speech-to-text model it says averages a 4.0% word error rate streaming and 2.6% non-streaming.

Both figures are attributed to Artificial Analysis in Google's own announcement, and both cover the same model in two different serving modes. That split matters more than the headline number, because the two modes ship as two separate APIs with two separate model IDs.

What actually shipped

Gemini 3.5 Transcribe is available in two forms, per Google's post:

• Real-time streaming — continuous, bidirectional streaming with sub-second latency via the Live API, using gemini-3.5-transcribe-live.
• Pre-recorded audio — meetings, call logs and similar, with speaker attribution and word-level timestamps, via the Interactions API using gemini-3.5-transcribe.

Google presents the model as a major advancement over Chirp 3, its previous transcription model, and frames the release around voice agents, real-time captioning and post-call analytics pipelines rather than bulk file transcription.

The numbers, and who measured them

Two benchmark sets appear in the announcement. The headline WER figures — 4.0% streaming and 2.6% non-streaming — are averages measured by Artificial Analysis. On the FLEURS multilingual benchmark, across what Google describes as a set of top languages and locales, the model reports 5.50% WER in streaming mode and 5.04% non-streaming.

The latency claim is the one worth reading twice. Google says time to final transcription improves by 70% over Chirp 3, again as measured by Artificial Analysis. That is a relative improvement against the company's own prior model, not an absolute latency figure, and no millisecond number is given. For a voice agent, time to final transcription is the number that decides whether a turn feels interactive — so the direction is useful, but you will need your own measurement on your own audio before sizing a pipeline around it.

The features that are not accuracy claims

Several capabilities in the release are about output shape rather than raw transcription quality:

• Disfluency cleanup. The model handles self-corrections — Google's own example is "let's meet Tuesday—no, Wednesday" — removes filler words, and auto-formats text.
• Custom vocabulary. You supply the terms; the model adapts to specialized jargon and unusual spellings instead of leaving you to fix them by hand afterwards.
• Language coverage. Automatic detection and transcription of over 85 languages, with regional accents and dialects.
• Speaker attribution. Accurate for up to three speakers in pre-recorded audio, with timestamps. Google labels support beyond three speakers as experimental.
• Function calling. The model can hand complex tasks such as image generation and file analysis to other Gemini models. This is currently limited to the Gemini macOS app.

The three-speaker ceiling is the practical constraint here. Diarization on a four-person meeting recording falls into the experimental bucket, which rules the pre-recorded path out of a lot of conference-call analytics work for now.

One model, not three

Coverage of this release needs a correction attached. The Verge reported that Google had mentioned two additional Gemini Audio models, 3.5 Live and 3.5 Live Experimental, in information provided before publication, then said afterwards that only 3.5 Transcribe was being announced. The Verge appended that correction on August 26.

So if you saw a three-model Gemini Audio refresh in early coverage, only one model is confirmed shipped. The Verge also notes that 3.5 Transcribe follows the launch of 3.5 Live Translate, and that Gemini 3.5 Pro — which the publication says Google promised to roll out in June — has still not been released.

Where you can use it today

Availability is split by audience, and most of it is preview:

• Developers — public preview in the Gemini API via Google AI Studio, and in Google Antigravity. AI Studio's Build mode lets you use the model to code by voice.
• Enterprises — public preview via the Gemini Enterprise Agent Platform, with Gemini Enterprise for Customer Experience listed as coming soon.
• Everyone — the Gemini app on macOS, in English; the new Rambler dictation feature on Gboard on Android, in select countries and languages; Chrome support is listed as coming soon.

Google also names developer platforms that expose the model through the Gemini Live API — Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel and Vision Agents — which is the fastest path in if you would rather not manage real-time media streaming yourself.

Is it worth switching?

If you are already on Chirp 3 inside Google's stack, the case is straightforward: same vendor, better reported WER, and a 70%-better time to final transcription by Google's measurement. The migration cost is an API change, since the streaming and pre-recorded paths now live behind different endpoints.

If you are evaluating from outside, three things are missing from the announcement. There is no pricing in it, so cost per audio hour is not comparable yet. Every accuracy and latency figure quoted is Google citing Artificial Analysis, not an independent head-to-head against non-Google models. And public preview is not a support commitment — the surfaces most people would deploy against are explicitly preview or "coming soon".

The sensible test is narrow and cheap. Take an hour of your worst audio — the noisy calls, the accented speakers, the recordings full of order IDs and postal codes, which Google specifically claims the model handles well — and run it through both paths. Measure WER yourself and time the final transcription. Then decide.

Two things to watch: whether pricing lands before general availability, and whether speaker attribution past three voices moves out of experimental. The second one gates more real workloads than the WER numbers do.